I did some analysis a couple of weeks ago, and most of the slow down in hived API response time isn't really related to rocksdb itself. The first issue was related to locking, but we improved that a lot already.
In my limited tests, there were usually two remaining causes for slow API call response times: 1) computation of results that could/should have been cached and 2) time to serialize output data (often supplying data that wasn't even strictly needed for most applications).
But the schema changes required to fix some of these issues will be much easier to do in postgresql. And I also see some nice potential benefits with postgresql having built-in interpreters (more about that later).